Scholarlink

This paper introduces an efficient task allocation paradigm for energy-aware deep learning inference in hybrid platforms that combine Field-Programmable Gate Arrays (FPGAs) and Graphics Processing Units (GPUs). Fulfilling the critical need of optimized computing for edge and cloud contexts, the paradigm dynamically allocates inference tasks to FPGAs or GPUs depending on energy efficiency and latency needs. It employs a prediction model that investigates properties of deep learning models—e.g., layer types, parameter numbers, and computational intensities—to project energy consumption and latency for all devices. The system then schedules tasks to the device that minimizes energy use while meeting user-specified latency constraints. Simulated experiments were performed over a range of deep learning models, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformers, to experimentally confirm the effectiveness of the framework. Experimental results indicate that the dynamic allocation strategy achieves up to 30% energy savings compared to static FPGA-only or GPU-only schemes without compromising inference latency. This work contributes to the Global progress towards achieving Energy Efficiency (UN SDG 7.3) for producing sustainable economic growth. By leveraging the strengths of FPGAs and GPUs in combination, this framework advances the frontiers of energy-aware computing through a scalable answer to the growing requirements for deep learning inference in both high-performance computing and resource-constrained settings.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dynamic Task Allocation Framework for Energy-Efficient Deep Learning Inference Using FPGA-GPU Hybrid Systems

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • B. Abirami,
  • V. Vasudevan

摘要

This paper introduces an efficient task allocation paradigm for energy-aware deep learning inference in hybrid platforms that combine Field-Programmable Gate Arrays (FPGAs) and Graphics Processing Units (GPUs). Fulfilling the critical need of optimized computing for edge and cloud contexts, the paradigm dynamically allocates inference tasks to FPGAs or GPUs depending on energy efficiency and latency needs. It employs a prediction model that investigates properties of deep learning models—e.g., layer types, parameter numbers, and computational intensities—to project energy consumption and latency for all devices. The system then schedules tasks to the device that minimizes energy use while meeting user-specified latency constraints. Simulated experiments were performed over a range of deep learning models, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformers, to experimentally confirm the effectiveness of the framework. Experimental results indicate that the dynamic allocation strategy achieves up to 30% energy savings compared to static FPGA-only or GPU-only schemes without compromising inference latency. This work contributes to the Global progress towards achieving Energy Efficiency (UN SDG 7.3) for producing sustainable economic growth. By leveraging the strengths of FPGAs and GPUs in combination, this framework advances the frontiers of energy-aware computing through a scalable answer to the growing requirements for deep learning inference in both high-performance computing and resource-constrained settings.